References

Tensorflow Official Benchmarks (May 2017, GitHub source): https://www.tensorflow.org/performance/benchmarks

IBM Power9 benchmark results (Nov 2017, 1.4.0): https://developer.ibm.com/linuxonpower/perfcol/perfcol-mldl/

Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour, Facebook (Jun 2017): https://research.fb.com/wp-content/uploads/2017/06/imagenet1kin1h5.pdf

Benchmark Source Code

https://github-dev.cs.illinois.edu/kindrtnk/DL

Official TF Benchmark System Characteristics

Instance type: NVIDIA® DGX-1™
GPU: 8x NVIDIA® Tesla® P100
OS: Ubuntu 16.04 LTS with tests run via Docker
CUDA / cuDNN: 8.0 / 5.1
TensorFlow GitHub hash: b1e174e
Benchmark GitHub hash: 9165a70
Build Command: bazel build -c opt --copt=-march="haswell" --config=cuda //tensorflow/tools/pip_package:build_pip_package
Disk: Local SSD
DataSet: ImageNet
Test Date: May 2017

Our System Characteristics (more details in GitHub Repo)

Instance type: IBM Power9, 8335-GTG AC922 server
CPU: 2x 20-core IBM POWER9 CPU @ 2.00GHz
SDRAM: 512G DDR4
GPU: 4x NVIDIA® Tesla® V100, 5120 cores, 16 GB HBM 2
Disk: Local SSD
OS: Red Hat Enterprise Linux Server release 7.4
Python Distribution: Anaconda python 3.6.2
CUDA / cuDNN: 9.1/7.0.5
TensorFLow Version: 1.5.0
DataSet: ImageNet (synthetic)
Precision: floating point 32 and 16
Test Date: Mar 25 2018

The following table is the result of running with the same configurations as the official Tensorflow benchmark mentioned in "Reference" section above:

	base_model	FP_type	batch_size	n_gpus	variable_update	local_parameter_device	image_per_sec
0	alexnet	16	512	1	replicated	N/A	8092.12
1	alexnet	16	512	2	replicated	N/A	15289.09
2	alexnet	16	512	4	replicated	N/A	25295.36
3	alexnet	32	512	1	replicated	N/A	4692.38
4	alexnet	32	512	2	replicated	N/A	8993.17
5	alexnet	32	512	4	replicated	N/A	15868.53
6	inception3	16	64	1	parameter_server	cpu_parameterDevice	423.90
7	inception3	16	64	2	parameter_server	cpu_parameterDevice	823.06
8	inception3	16	64	4	parameter_server	cpu_parameterDevice	1561.19
9	inception3	32	64	1	parameter_server	cpu_parameterDevice	245.36
10	inception3	32	64	2	parameter_server	cpu_parameterDevice	472.07
11	inception3	32	64	4	parameter_server	cpu_parameterDevice	918.71
12	resnet152	16	64	1	parameter_server	cpu_parameterDevice	283.59
13	resnet152	16	64	2	parameter_server	cpu_parameterDevice	531.70
14	resnet152	16	64	4	parameter_server	cpu_parameterDevice	975.86
15	resnet152	32	64	1	parameter_server	cpu_parameterDevice	148.61
16	resnet152	32	64	2	parameter_server	cpu_parameterDevice	287.02
17	resnet152	32	64	4	parameter_server	cpu_parameterDevice	567.19
18	resnet50	16	64	1	parameter_server	cpu_parameterDevice	618.52
19	resnet50	16	64	2	parameter_server	cpu_parameterDevice	1174.06
20	resnet50	16	64	4	parameter_server	cpu_parameterDevice	2294.24
21	resnet50	32	64	1	parameter_server	cpu_parameterDevice	364.54
22	resnet50	32	64	2	parameter_server	cpu_parameterDevice	719.03
23	resnet50	32	64	4	parameter_server	cpu_parameterDevice	1402.16
24	vgg16	16	64	1	replicated	N/A	405.58
25	vgg16	16	64	2	replicated	N/A	766.43
26	vgg16	16	64	4	replicated	N/A	1264.35
27	vgg16	32	64	1	replicated	N/A	234.22
28	vgg16	32	64	2	replicated	N/A	451.64
29	vgg16	32	64	4	replicated	N/A	809.41

This figure compares the result we get with Tensorflow official ones.

Green bars stand for our benchmark results using floating point 16.

Red bars are the official Tensorflow result.

Blue bars stand for our benchmark results using floating point 32.

This figure shows the performance ratio of our floating point 16 and 32 benchmarks with respect to Tensorflow official results:

The following table provides a more comprehensive benchmark results on our system:

	base_model	FP_type	batch_size	n_gpus	variable_update	local_parameter_device	image_per_sec
0	alexnet	16	512	1	parameter_server	N/A	8176.13
1	alexnet	16	512	1	replicated	N/A	8092.12
2	alexnet	16	512	2	parameter_server	N/A	15724.07
3	alexnet	16	512	2	replicated	N/A	15289.09
4	alexnet	16	512	4	parameter_server	N/A	26709.08
5	alexnet	16	512	4	replicated	N/A	25295.36
6	alexnet	32	512	1	parameter_server	N/A	4645.89
7	alexnet	32	512	1	replicated	N/A	4692.38
8	alexnet	32	512	2	parameter_server	N/A	8994.28
9	alexnet	32	512	2	replicated	N/A	8993.17
10	alexnet	32	512	4	parameter_server	N/A	15563.43
11	alexnet	32	512	4	replicated	N/A	15868.53
12	inception3	16	32	1	parameter_server	gpu_parameterDevice	368.16
13	inception3	16	32	1	parameter_server	cpu_parameterDevice	289.93
14	inception3	16	32	1	replicated	cpu_parameterDevice	345.22
15	inception3	16	32	1	replicated	gpu_parameterDevice	336.22
16	inception3	16	32	2	parameter_server	gpu_parameterDevice	569.06
17	inception3	16	32	2	parameter_server	cpu_parameterDevice	593.43
18	inception3	16	32	2	replicated	cpu_parameterDevice	610.20
19	inception3	16	32	2	replicated	gpu_parameterDevice	594.11
20	inception3	16	32	4	parameter_server	gpu_parameterDevice	972.40
21	inception3	16	32	4	parameter_server	cpu_parameterDevice	1030.18
22	inception3	16	32	4	replicated	gpu_parameterDevice	1076.02
23	inception3	16	32	4	replicated	cpu_parameterDevice	1064.68
24	inception3	16	64	1	parameter_server	cpu_parameterDevice	423.90
25	inception3	16	64	1	parameter_server	gpu_parameterDevice	440.50
26	inception3	16	64	1	replicated	gpu_parameterDevice	436.09
27	inception3	16	64	1	replicated	cpu_parameterDevice	436.16
28	inception3	16	64	2	parameter_server	cpu_parameterDevice	823.06
29	inception3	16	64	2	parameter_server	gpu_parameterDevice	811.04
30	inception3	16	64	2	replicated	cpu_parameterDevice	842.72
31	inception3	16	64	2	replicated	gpu_parameterDevice	848.60
32	inception3	16	64	4	parameter_server	cpu_parameterDevice	1561.19
33	inception3	16	64	4	parameter_server	gpu_parameterDevice	1502.41
34	inception3	16	64	4	replicated	cpu_parameterDevice	1701.83
35	inception3	16	64	4	replicated	gpu_parameterDevice	1605.17
36	inception3	16	128	1	replicated	no_parameterDevice	491.12
37	inception3	16	128	2	replicated	no_parameterDevice	972.62
38	inception3	16	128	4	replicated	no_parameterDevice	1926.57
39	inception3	16	256	1	replicated	no_parameterDevice	521.54
40	inception3	16	256	2	replicated	no_parameterDevice	1032.37
41	inception3	16	256	4	replicated	no_parameterDevice	2043.78
42	inception3	32	32	1	parameter_server	gpu_parameterDevice	224.14
43	inception3	32	32	1	parameter_server	cpu_parameterDevice	217.93
44	inception3	32	32	1	replicated	cpu_parameterDevice	224.14
45	inception3	32	32	1	replicated	gpu_parameterDevice	225.85
46	inception3	32	32	2	parameter_server	gpu_parameterDevice	414.47
47	inception3	32	32	2	parameter_server	cpu_parameterDevice	424.24
48	inception3	32	32	2	replicated	cpu_parameterDevice	431.49
49	inception3	32	32	2	replicated	gpu_parameterDevice	439.14
50	inception3	32	32	4	parameter_server	cpu_parameterDevice	788.87
51	inception3	32	32	4	parameter_server	gpu_parameterDevice	747.39
52	inception3	32	32	4	replicated	cpu_parameterDevice	856.59
53	inception3	32	32	4	replicated	gpu_parameterDevice	829.10
54	inception3	32	64	1	parameter_server	gpu_parameterDevice	247.98
55	inception3	32	64	1	parameter_server	cpu_parameterDevice	245.36
56	inception3	32	64	1	replicated	cpu_parameterDevice	247.98
57	inception3	32	64	1	replicated	gpu_parameterDevice	247.96
58	inception3	32	64	2	parameter_server	cpu_parameterDevice	472.07
59	inception3	32	64	2	parameter_server	gpu_parameterDevice	451.69
60	inception3	32	64	2	replicated	gpu_parameterDevice	491.18
61	inception3	32	64	2	replicated	cpu_parameterDevice	486.19
62	inception3	32	64	4	parameter_server	gpu_parameterDevice	919.57
63	inception3	32	64	4	parameter_server	cpu_parameterDevice	918.71
64	inception3	32	64	4	replicated	gpu_parameterDevice	963.25
65	inception3	32	64	4	replicated	cpu_parameterDevice	962.94
66	resnet152	16	32	1	parameter_server	cpu_parameterDevice	201.65
67	resnet152	16	32	1	parameter_server	gpu_parameterDevice	229.96
68	resnet152	16	32	1	replicated	gpu_parameterDevice	218.05
69	resnet152	16	32	1	replicated	cpu_parameterDevice	193.48
70	resnet152	16	32	2	parameter_server	cpu_parameterDevice	332.26
71	resnet152	16	32	2	parameter_server	gpu_parameterDevice	341.70
72	resnet152	16	32	2	replicated	cpu_parameterDevice	329.15
73	resnet152	16	32	2	replicated	gpu_parameterDevice	375.69
74	resnet152	16	32	4	parameter_server	gpu_parameterDevice	537.87
75	resnet152	16	32	4	parameter_server	cpu_parameterDevice	593.27
76	resnet152	16	32	4	replicated	cpu_parameterDevice	664.84
77	resnet152	16	32	4	replicated	gpu_parameterDevice	624.26
78	resnet152	16	64	1	parameter_server	gpu_parameterDevice	290.69
79	resnet152	16	64	1	parameter_server	cpu_parameterDevice	283.59
80	resnet152	16	64	1	replicated	cpu_parameterDevice	287.45
81	resnet152	16	64	1	replicated	gpu_parameterDevice	287.39
82	resnet152	16	64	2	parameter_server	cpu_parameterDevice	531.70
83	resnet152	16	64	2	parameter_server	gpu_parameterDevice	510.79
84	resnet152	16	64	2	replicated	cpu_parameterDevice	562.41
85	resnet152	16	64	2	replicated	gpu_parameterDevice	574.11
86	resnet152	16	64	4	parameter_server	cpu_parameterDevice	975.86
87	resnet152	16	64	4	parameter_server	gpu_parameterDevice	894.71
88	resnet152	16	64	4	replicated	cpu_parameterDevice	1075.84
89	resnet152	16	64	4	replicated	gpu_parameterDevice	945.07
90	resnet152	16	128	1	replicated	no_parameterDevice	330.65
91	resnet152	16	128	2	replicated	no_parameterDevice	648.43
92	resnet152	16	128	4	replicated	no_parameterDevice	1288.11
93	resnet152	32	32	1	parameter_server	cpu_parameterDevice	131.59
94	resnet152	32	32	1	parameter_server	gpu_parameterDevice	137.46
95	resnet152	32	32	1	replicated	gpu_parameterDevice	137.48
96	resnet152	32	32	1	replicated	cpu_parameterDevice	137.46
97	resnet152	32	32	2	parameter_server	cpu_parameterDevice	252.06
98	resnet152	32	32	2	parameter_server	gpu_parameterDevice	258.02
99	resnet152	32	32	2	replicated	cpu_parameterDevice	266.06
100	resnet152	32	32	2	replicated	gpu_parameterDevice	269.09
101	resnet152	32	32	4	parameter_server	cpu_parameterDevice	475.65
102	resnet152	32	32	4	parameter_server	gpu_parameterDevice	428.43
103	resnet152	32	32	4	replicated	cpu_parameterDevice	531.89
104	resnet152	32	32	4	replicated	gpu_parameterDevice	510.74
105	resnet152	32	64	1	parameter_server	cpu_parameterDevice	148.61
106	resnet152	32	64	1	parameter_server	gpu_parameterDevice	153.30
107	resnet152	32	64	1	replicated	cpu_parameterDevice	152.38
108	resnet152	32	64	1	replicated	gpu_parameterDevice	153.30
109	resnet152	32	64	2	parameter_server	gpu_parameterDevice	297.53
110	resnet152	32	64	2	parameter_server	cpu_parameterDevice	287.02
111	resnet152	32	64	2	replicated	gpu_parameterDevice	304.75
112	resnet152	32	64	2	replicated	cpu_parameterDevice	302.79
113	resnet152	32	64	4	parameter_server	cpu_parameterDevice	567.19
114	resnet152	32	64	4	parameter_server	gpu_parameterDevice	546.67
115	resnet152	32	64	4	replicated	cpu_parameterDevice	591.06
116	resnet152	32	64	4	replicated	gpu_parameterDevice	587.99
117	resnet152_v2	16	32	1	parameter_server	cpu_parameterDevice	210.39
118	resnet152_v2	16	32	1	parameter_server	gpu_parameterDevice	234.22
119	resnet152_v2	16	32	1	replicated	cpu_parameterDevice	218.01
120	resnet152_v2	16	32	1	replicated	gpu_parameterDevice	225.86
121	resnet152_v2	16	32	2	parameter_server	cpu_parameterDevice	319.55
122	resnet152_v2	16	32	2	parameter_server	gpu_parameterDevice	357.69
123	resnet152_v2	16	32	2	replicated	gpu_parameterDevice	377.47
124	resnet152_v2	16	32	2	replicated	cpu_parameterDevice	384.85
125	resnet152_v2	16	32	4	parameter_server	cpu_parameterDevice	652.52
126	resnet152_v2	16	32	4	parameter_server	gpu_parameterDevice	550.14
127	resnet152_v2	16	32	4	replicated	cpu_parameterDevice	673.32
128	resnet152_v2	16	32	4	replicated	gpu_parameterDevice	632.01
129	resnet152_v2	16	64	1	parameter_server	gpu_parameterDevice	294.10
130	resnet152_v2	16	64	1	parameter_server	cpu_parameterDevice	290.34
131	resnet152_v2	16	64	1	replicated	gpu_parameterDevice	294.56
132	resnet152_v2	16	64	1	replicated	cpu_parameterDevice	294.12
133	resnet152_v2	16	64	2	parameter_server	gpu_parameterDevice	532.29
134	resnet152_v2	16	64	2	parameter_server	cpu_parameterDevice	542.82
135	resnet152_v2	16	64	2	replicated	gpu_parameterDevice	575.56
136	resnet152_v2	16	64	2	replicated	cpu_parameterDevice	574.35
137	resnet152_v2	16	64	4	parameter_server	cpu_parameterDevice	1030.82
138	resnet152_v2	16	64	4	parameter_server	gpu_parameterDevice	944.88
139	resnet152_v2	16	64	4	replicated	gpu_parameterDevice	1032.10
140	resnet152_v2	16	64	4	replicated	cpu_parameterDevice	1159.66
141	resnet152_v2	16	128	1	replicated	no_parameterDevice	335.02
142	resnet152_v2	16	128	2	replicated	no_parameterDevice	661.07
143	resnet152_v2	16	128	4	replicated	no_parameterDevice	1296.60
144	resnet152_v2	32	32	1	parameter_server	cpu_parameterDevice	132.99
145	resnet152_v2	32	32	1	parameter_server	gpu_parameterDevice	139.09
146	resnet152_v2	32	32	1	replicated	gpu_parameterDevice	140.51
147	resnet152_v2	32	32	1	replicated	cpu_parameterDevice	138.95
148	resnet152_v2	32	32	2	parameter_server	gpu_parameterDevice	252.95
149	resnet152_v2	32	32	2	parameter_server	cpu_parameterDevice	257.58
150	resnet152_v2	32	32	2	replicated	gpu_parameterDevice	271.95
151	resnet152_v2	32	32	2	replicated	cpu_parameterDevice	271.85
152	resnet152_v2	32	32	4	parameter_server	gpu_parameterDevice	447.54
153	resnet152_v2	32	32	4	parameter_server	cpu_parameterDevice	484.61
154	resnet152_v2	32	32	4	replicated	gpu_parameterDevice	532.23
155	resnet152_v2	32	32	4	replicated	cpu_parameterDevice	532.04
156	resnet152_v2	32	64	1	parameter_server	gpu_parameterDevice	155.19
157	resnet152_v2	32	64	1	parameter_server	cpu_parameterDevice	151.34
158	resnet152_v2	32	64	1	replicated	gpu_parameterDevice	154.24
159	resnet152_v2	32	64	1	replicated	cpu_parameterDevice	154.24
160	resnet152_v2	32	64	2	parameter_server	cpu_parameterDevice	298.92
161	resnet152_v2	32	64	2	parameter_server	gpu_parameterDevice	295.79
162	resnet152_v2	32	64	2	replicated	gpu_parameterDevice	308.46
163	resnet152_v2	32	64	2	replicated	cpu_parameterDevice	306.41
164	resnet152_v2	32	64	4	parameter_server	cpu_parameterDevice	557.58
165	resnet152_v2	32	64	4	parameter_server	gpu_parameterDevice	526.94
166	resnet152_v2	32	64	4	replicated	cpu_parameterDevice	601.99
167	resnet152_v2	32	64	4	replicated	gpu_parameterDevice	595.10
168	resnet50	16	32	1	parameter_server	cpu_parameterDevice	486.00
169	resnet50	16	32	1	parameter_server	gpu_parameterDevice	521.21
170	resnet50	16	32	1	replicated	cpu_parameterDevice	471.64
171	resnet50	16	32	1	replicated	gpu_parameterDevice	516.20
172	resnet50	16	32	2	parameter_server	gpu_parameterDevice	824.17
173	resnet50	16	32	2	parameter_server	cpu_parameterDevice	891.88
174	resnet50	16	32	2	replicated	cpu_parameterDevice	981.99
175	resnet50	16	32	2	replicated	gpu_parameterDevice	953.31
176	resnet50	16	32	4	parameter_server	cpu_parameterDevice	1627.05
177	resnet50	16	32	4	parameter_server	gpu_parameterDevice	1502.83
178	resnet50	16	32	4	replicated	cpu_parameterDevice	1834.62
179	resnet50	16	32	4	replicated	gpu_parameterDevice	1598.16
180	resnet50	16	64	1	parameter_server	cpu_parameterDevice	618.52
181	resnet50	16	64	1	parameter_server	gpu_parameterDevice	641.96
182	resnet50	16	64	1	replicated	cpu_parameterDevice	632.34
183	resnet50	16	64	1	replicated	gpu_parameterDevice	638.60
184	resnet50	16	64	2	parameter_server	cpu_parameterDevice	1174.06
185	resnet50	16	64	2	parameter_server	gpu_parameterDevice	1221.21
186	resnet50	16	64	2	replicated	cpu_parameterDevice	1245.83
187	resnet50	16	64	2	replicated	gpu_parameterDevice	1239.34
188	resnet50	16	64	4	parameter_server	cpu_parameterDevice	2294.24
189	resnet50	16	64	4	parameter_server	gpu_parameterDevice	2199.00
190	resnet50	16	64	4	replicated	cpu_parameterDevice	2489.61
191	resnet50	16	64	4	replicated	gpu_parameterDevice	2376.21
192	resnet50	16	128	1	replicated	no_parameterDevice	712.54
193	resnet50	16	128	2	replicated	no_parameterDevice	1419.40
194	resnet50	16	128	4	replicated	no_parameterDevice	2733.84
195	resnet50	16	256	1	replicated	no_parameterDevice	749.41
196	resnet50	16	256	2	replicated	no_parameterDevice	1477.06
197	resnet50	16	256	4	replicated	no_parameterDevice	2932.09
198	resnet50	32	32	1	parameter_server	cpu_parameterDevice	324.01
199	resnet50	32	32	1	parameter_server	gpu_parameterDevice	336.18
200	resnet50	32	32	1	replicated	cpu_parameterDevice	332.84
201	resnet50	32	32	1	replicated	gpu_parameterDevice	332.86
202	resnet50	32	32	2	parameter_server	gpu_parameterDevice	642.03
203	resnet50	32	32	2	parameter_server	cpu_parameterDevice	621.98
204	resnet50	32	32	2	replicated	cpu_parameterDevice	658.53
205	resnet50	32	32	2	replicated	gpu_parameterDevice	648.60
206	resnet50	32	32	4	parameter_server	gpu_parameterDevice	1053.87
207	resnet50	32	32	4	parameter_server	cpu_parameterDevice	1201.09
208	resnet50	32	32	4	replicated	gpu_parameterDevice	1246.13
209	resnet50	32	32	4	replicated	cpu_parameterDevice	1282.85
210	resnet50	32	64	1	parameter_server	cpu_parameterDevice	364.54
211	resnet50	32	64	1	parameter_server	gpu_parameterDevice	371.99
212	resnet50	32	64	1	replicated	cpu_parameterDevice	371.97
213	resnet50	32	64	1	replicated	gpu_parameterDevice	372.02
214	resnet50	32	64	2	parameter_server	cpu_parameterDevice	719.03
215	resnet50	32	64	2	parameter_server	gpu_parameterDevice	722.70
216	resnet50	32	64	2	replicated	cpu_parameterDevice	722.34
217	resnet50	32	64	2	replicated	gpu_parameterDevice	736.44
218	resnet50	32	64	4	parameter_server	gpu_parameterDevice	1317.22
219	resnet50	32	64	4	parameter_server	cpu_parameterDevice	1402.16
220	resnet50	32	64	4	replicated	cpu_parameterDevice	1424.48
221	resnet50	32	64	4	replicated	gpu_parameterDevice	1424.51
222	resnet50_v2	16	32	1	parameter_server	cpu_parameterDevice	510.22
223	resnet50_v2	16	32	1	parameter_server	gpu_parameterDevice	537.68
224	resnet50_v2	16	32	1	replicated	gpu_parameterDevice	505.88
225	resnet50_v2	16	32	1	replicated	cpu_parameterDevice	516.14
226	resnet50_v2	16	32	2	parameter_server	gpu_parameterDevice	953.22
227	resnet50_v2	16	32	2	parameter_server	cpu_parameterDevice	942.64
228	resnet50_v2	16	32	2	replicated	gpu_parameterDevice	1011.77
229	resnet50_v2	16	32	2	replicated	cpu_parameterDevice	1000.48
230	resnet50_v2	16	32	4	parameter_server	cpu_parameterDevice	1643.04
231	resnet50_v2	16	32	4	parameter_server	gpu_parameterDevice	1517.55
232	resnet50_v2	16	32	4	replicated	gpu_parameterDevice	1703.01
233	resnet50_v2	16	32	4	replicated	cpu_parameterDevice	1778.21
234	resnet50_v2	16	64	1	parameter_server	cpu_parameterDevice	647.49
235	resnet50_v2	16	64	1	parameter_server	gpu_parameterDevice	655.20
236	resnet50_v2	16	64	1	replicated	gpu_parameterDevice	648.63
237	resnet50_v2	16	64	1	replicated	cpu_parameterDevice	655.02
238	resnet50_v2	16	64	2	parameter_server	cpu_parameterDevice	1232.88
239	resnet50_v2	16	64	2	parameter_server	gpu_parameterDevice	1200.12
240	resnet50_v2	16	64	2	replicated	cpu_parameterDevice	1283.14
241	resnet50_v2	16	64	2	replicated	gpu_parameterDevice	1297.01
242	resnet50_v2	16	64	4	parameter_server	gpu_parameterDevice	2176.56
243	resnet50_v2	16	64	4	parameter_server	cpu_parameterDevice	2347.95
244	resnet50_v2	16	64	4	replicated	cpu_parameterDevice	2553.59
245	resnet50_v2	16	64	4	replicated	gpu_parameterDevice	2492.30
246	resnet50_v2	16	128	1	replicated	no_parameterDevice	733.20
247	resnet50_v2	16	128	2	replicated	no_parameterDevice	1445.30
248	resnet50_v2	16	128	4	replicated	no_parameterDevice	2771.44
249	resnet50_v2	16	256	1	replicated	no_parameterDevice	766.46
250	resnet50_v2	16	256	2	replicated	no_parameterDevice	1532.91
251	resnet50_v2	16	256	4	replicated	no_parameterDevice	2997.32
252	resnet50_v2	32	32	1	parameter_server	cpu_parameterDevice	332.66
253	resnet50_v2	32	32	1	parameter_server	gpu_parameterDevice	347.81
254	resnet50_v2	32	32	1	replicated	cpu_parameterDevice	344.23
255	resnet50_v2	32	32	1	replicated	gpu_parameterDevice	345.27
256	resnet50_v2	32	32	2	parameter_server	gpu_parameterDevice	672.42
257	resnet50_v2	32	32	2	parameter_server	cpu_parameterDevice	650.97
258	resnet50_v2	32	32	2	replicated	cpu_parameterDevice	665.32
259	resnet50_v2	32	32	2	replicated	gpu_parameterDevice	683.57
260	resnet50_v2	32	32	4	parameter_server	cpu_parameterDevice	1248.53
261	resnet50_v2	32	32	4	parameter_server	gpu_parameterDevice	1101.25
262	resnet50_v2	32	32	4	replicated	gpu_parameterDevice	1258.33
263	resnet50_v2	32	32	4	replicated	cpu_parameterDevice	1303.18
264	resnet50_v2	32	64	1	parameter_server	cpu_parameterDevice	371.81
265	resnet50_v2	32	64	1	parameter_server	gpu_parameterDevice	383.19
266	resnet50_v2	32	64	1	replicated	gpu_parameterDevice	377.57
267	resnet50_v2	32	64	1	replicated	cpu_parameterDevice	383.30
268	resnet50_v2	32	64	2	parameter_server	cpu_parameterDevice	732.40
269	resnet50_v2	32	64	2	parameter_server	gpu_parameterDevice	747.31
270	resnet50_v2	32	64	2	replicated	gpu_parameterDevice	751.50
271	resnet50_v2	32	64	2	replicated	cpu_parameterDevice	751.11
272	resnet50_v2	32	64	4	parameter_server	gpu_parameterDevice	1353.14
273	resnet50_v2	32	64	4	parameter_server	cpu_parameterDevice	1423.04
274	resnet50_v2	32	64	4	replicated	gpu_parameterDevice	1445.11
275	resnet50_v2	32	64	4	replicated	cpu_parameterDevice	1444.34
276	vgg16	16	32	1	parameter_server	N/A	379.42
277	vgg16	16	32	1	replicated	N/A	379.52
278	vgg16	16	32	2	parameter_server	N/A	722.64
279	vgg16	16	32	2	replicated	N/A	676.75
280	vgg16	16	32	4	parameter_server	N/A	991.95
281	vgg16	16	32	4	replicated	N/A	1001.89
282	vgg16	16	64	1	parameter_server	N/A	403.87
283	vgg16	16	64	1	replicated	N/A	405.58
284	vgg16	16	64	2	parameter_server	N/A	778.24
285	vgg16	16	64	2	replicated	N/A	766.43
286	vgg16	16	64	4	parameter_server	N/A	1264.47
287	vgg16	16	64	4	replicated	N/A	1264.35
288	vgg16	16	128	1	replicated	N/A	425.12
289	vgg16	16	128	2	replicated	N/A	822.49
290	vgg16	16	128	4	replicated	N/A	1466.10
291	vgg16	16	256	1	replicated	N/A	396.79
292	vgg16	16	256	2	replicated	N/A	778.23
293	vgg16	16	256	4	replicated	N/A	1371.82
294	vgg16	32	32	1	parameter_server	N/A	225.86
295	vgg16	32	32	1	replicated	N/A	225.86
296	vgg16	32	32	2	parameter_server	N/A	424.41
297	vgg16	32	32	2	replicated	N/A	418.82
298	vgg16	32	32	4	parameter_server	N/A	692.95
299	vgg16	32	32	4	replicated	N/A	683.66
300	vgg16	32	64	1	parameter_server	N/A	236.39
301	vgg16	32	64	1	replicated	N/A	234.22
302	vgg16	32	64	2	parameter_server	N/A	455.67
303	vgg16	32	64	2	replicated	N/A	451.64
304	vgg16	32	64	4	parameter_server	N/A	815.81
305	vgg16	32	64	4	replicated	N/A	809.41

Child pages

References

Benchmark Source Code

Official TF Benchmark System Characteristics

Our System Characteristics (more details in GitHub Repo)

The following table is the result of running with the same configurations as the official Tensorflow benchmark mentioned in "Reference" section above:

This figure compares the result we get with Tensorflow official ones.

This figure shows the performance ratio of our floating point 16 and 32 benchmarks with respect to Tensorflow official results:

The following table provides a more comprehensive benchmark results on our system:

POWER8 (p8)

Child pages

Benchmark Mar_26_2018

References

Benchmark Source Code

Official TF Benchmark System Characteristics

Our System Characteristics (more details in GitHub Repo)

The following table is the result of running with the same configurations as the official Tensorflow benchmark mentioned in "Reference" section above:

This figure compares the result we get with Tensorflow official ones.

This figure shows the performance ratio of our floating point 16 and 32 benchmarks with respect to Tensorflow official results:

The following table provides a more comprehensive benchmark results on our system:

POWER8 (p8)